Author: Graham Laight
Date: 02:40:43 10/11/00
Go up one level in this thread
On October 10, 2000 at 20:56:58, Michael Neish wrote: > >Hello, > >I'd like to learn a little about pattern recognition as applied to Chess, and >would like to ask whether anyone could direct me to some information either >online or in the literature, or has first-hand experience at this sort of >programming (or knowledge), and would like to comment about it. > >All I've heard up to now is that techniques like genetic algorithms and neural >networks fail woefully at Chess. > >Well, I got tired of trying to etch the same path as everyone else is (null >move, mobility scores, hash tables ... ) and have decided to try something >different. Please forgive my naivety. :) > >Cheers, > >Mike. Good thinking, Mike! Whatever you do, don't let the people here discourage you. When you ask to discuss this kind of thing, you tend to get a lot of "can't be done", "not in our lifetime", and other such ill thought out remarks. Well the fact is that most people in the chess community have dedicated their lives to the fast search method, and are not willing to countenance the possibility of alternative ways. One famous science author (can't remember his name, but it's from his book "The red Queen") said that science does not usually advance until the people in charge, who believe in the prevailing doctrine, die. You know what I say? Of course NNs can recognise patterns - that's EXACTLY what they do well! Nobody has done it yet for chess - but if NN chess pattern recognition had received as much study as fast searching has, I think it would be doing it well by now. Genetic algorithms? Yes - with some imagination these could be used as well. If you make a selection of "components" for evaluating chess positions, a genetic algorithm might be able to both select the best combination of components and tune them to make the computer evaluate positions to the satisfaction of a strong player. My suggestion has always been to select among (a database of) different evaluation functions, depending on the current position. Whatever you do in the knowledge arena, you're going to pay a high price in terms of NPS. Human NPS is probably about 2 to 4. Given the persistent weaknesses of today's strongest programs, there could be rich rewards for the individual or team that addresses these issues. links to get you started: http://forum.swarthmore.edu/~jay/learn-game/systems/neurochess.html http://forum.swarthmore.edu/~jay/learn-game/systems/sal.html
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.